Integrated into Google Gemini but reserved for subscribers of the Ultra subscription, the new Gemini 2.5 Deep Think model is based on a unprecedented multi-agent approach, which allows it to “think” simultaneously with several things. It is the model that won a gold medal in mathematics a few days ago.
Pending GPT-5, whose announcement could take place next week, Google releases heavy artillery. In a press release published on August 1ᵉʳ, the web giant announces deploying Gemini 2.5 Deep Think, its most advanced model to date, with its Ultra subscribers (its package at 275 euros per month). It is a version derived from the model that won a mathematics gold medal in July 2025, the first capable of “thinking” at several things at the same time.
With Gemini 2.5 Deep Think, Google gets ahead of its competitors
After having widely caught up with Gemini 2.5 Pro, not to say exceed everyone, Google puts a layer with its new Deep Think model.
Your data deserve to be better protected.
Sociations, viruses and ransomware are no longer inevitable. Protect your privacy and personal data with Bitdefender and enjoy your digital life with confidence.
Presented at the Google I/O in May, Gemini 2.5 Deep Think was hitherto accessible only in an experimental form intended for reselected reselects by Google. Deep Think is available today in the Gemini application for ultra subscribers, in an unprecedented version. Its results are less precise than with the model tested during the international Mathematics Olympiads, but its answers are faster (a few minutes versus several hours). Gemini 2.5 Deep Think would have won a bronze medal with mathematics Olympiads according to Google.
Gemini 2.5 Deep Think is the best model of AI on the market? In fact, everything suggests that yes. The model has a “parallel” capacity for reflection.
Unlike other models of reflection, which automate several research, wonder and issue chain hypotheses before responding, Gemini 2.5 Deep Think generates several ideas simultaneously and can envisage them all at the same time. Each agent issues his own hypotheses and confronts them with others. At the end, Gemini 2.5 Deep Think, a bit like a jury, generates a consensual result based on all the tracks mentioned. This behavior allows him to merge ideas, change his mind and improve his skill in solving complex problems. His time of reflection is longer than in other modes like Deep Research, which allows Deep Think to work more hypotheses than other models on the market.
In view of its exceptional abilities, Deep Think is not intended for all users. The model can analyze data in an ultra-precise way and potentially generate complex code. It marks a first step towards planning models capable of managing problems that humans would struggle to solve, while allowing Google to overshadow its competitors in benchmarks.
Given the necessary calculation power, Gemini 2.5 Deep Think should remain reserved for ultra subscribers for a long time. Google also says it also deployed for a few partner mathematicians access to its gold medalist model. The timing of his announcement is certainly not trivial: Google knows that Optai, with GPT-5, could take the crown of the best model of AI.
All tech news in the blink of an eye
Add Numerama to your home screen and stay connected to the future!